Add docker protocol support for llama-server model loading

[ericcurtin]

To pull and run models via: llama-server -dr gemma3
Add some validators and sanitizers for Docker Model urls and metadata

[ericcurtin]

@CISC @ggerganov PTAL

[Copilot]

Pull Request Overview

This PR adds Docker registry support to llama-server, enabling users to pull and run AI models directly from Docker Hub using the docker:// protocol. The implementation handles Docker registry authentication, manifest parsing, and blob downloading to cache models locally.

• Adds Docker URL parsing and resolution functionality to download GGUF models from Docker registries
• Integrates Docker model resolution into the existing model loading pipeline
• Implements streaming download with proper authentication and caching support

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 6 comments.

File	Description
common/common.cpp	Integrates Docker model resolution into the model loading pipeline and updates error messages
common/arg.h	Adds function declaration for Docker model resolution
common/arg.cpp	Implements complete Docker registry functionality including authentication, manifest parsing, and blob downloading

[ericcurtin]

@JohannesGaessler @slaren PTAL

[ericcurtin]

@danbev PTAL

[ericcurtin]

@ggerganov @JohannesGaessler @slaren @danbev struggle to get this reviewed, if you guys have cycles I'd appreciate it.

[ericcurtin]

I'm also having a change of heart, thinking of changing to a:

-d/--docker-repo

option, at least the string would be the same then as suggested in Docker Hub and the one used in Docker Model Runner. It would also be more consistent with the huggingface argument approach.

[ericcurtin]

@ggerganov @danbev ready for re-review

[ericcurtin]

Added resumable downloads in a second commit, models can be large and redownloading models from scratch on interrupted connections can be a pain and a waste of bandwidth

[ericcurtin]

@ggerganov ready for re-review

[ggerganov]

The docker-related functions seem ok.

I'm not confident about the changes to common_download_file_single to support resumable downloads. Either wait for someone to review this part in details, or move it to a separate PR.

[ericcurtin]

The docker-related functions seem ok.

I'm not confident about the changes to common_download_file_single to support resumable downloads. Either wait for someone to review this part in details, or move it to a separate PR.

SGTM, I do think that resumable downloads bit is important in the next PR, whether it's huggingface, docker, etc. Somebody has to pay the cloud bill of all the wasted petabytes that are retransferred because of retries.

And of course having to start a download from the start again because of an interrupted connection simply being annoying.

Sometimes that can make larger models impossible to download for people.

Even some servers have server side timeouts if you down finish the download in a certain time.

Resumable downloads solves these things, client side.

[ericcurtin]

@ggerganov all done, it's just the docker pulling change now

[ericcurtin]

Getting build problems unrelated to this PR:

/Users/runner/work/llama.cpp/llama.cpp/ggml/src/ggml-blas/ggml-blas.cpp:143:13: error: 'cblas_sgemm' is only available on macOS 13.3 or newer [-Werror,-Wunguarded-availability-new]
            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
            ^~~~~~~~~~~

x86_64 macOS

Gonna try a rebuild.

[ericcurtin]

@ggerganov green build!